Asking machines to identify images

[Ryan_m_b]

I can't get over how cool/creepy this is. Researchers took image recognition neural networks and asked them to look for certain images in random noise and once identified modify the image to highlight the pattern. That new image is fed back into the machine and it's asked to do the same again. After multiple repetitions the pictures are bizarre, beautiful or both.

http://www.theguardian.com/technolo...twork-androids-dream-electric-sheep?CMP=fb_gu

 

[Hesch]

Researchers took image recognition neural networks and asked them to look for certain images in random noise and once identified modify the image to highlight the pattern.

I see what you mean ( by most of the images ). I would not call it "pattern recognition" but rather a LSD-trip.

But I think you know, that "real" pattern recognition uses mathematical rigid methods like Fourier-transforms, Hough-transforms, lens correcting functions, amongst other methods, to measure what is going on in an industrial production ?

 

[lisab]

What a cool idea!

 

[Ryan_m_b]

I see what you mean ( by most of the images ). I would not call it "pattern recognition" but rather a LSD-trip.

What makes it not pattern recognition? It's not necessarily good pattern recognition initially given that the machine is detecting patterns that aren't there but after it has edited them in slightly subsequent tests detect that patter.

But I think you know, that "real" pattern recognition uses mathematical rigid methods like Fourier-transforms, Hough-transforms, lens correcting functions, amongst other methods, to measure what is going on in an industrial production ?

Why would you think I know that and what makes any of that "real"?

 

[nsaspook]

Finding false patterns in noise something computers are really good at. Program them to find something and boom, they will with high confidence.

http://www.wired.com/2009/09/fmrisalmon/
http://www.edvul.com/pdf/VulHarrisWinkielmanPashler-PPS-2009.pdf

 

[ShayanJ]

That is cool. They can arrange an exhibition.

 

[Hesch]

Why would you think I know that and what makes any of that "real"?

Sorry, I thought you meant it as a joke.

I think that this example is real:

The original photo is "filled" with noise which has some distinct feature. You Fourier transform (FT) the noisy picture and will find this feature in the Fourier transform. Remove it from the transform and make the inverse Fourier transform (IFT). Then you'll get the picture to the right.

Say you have a photo of a car driving by. Due to the speed of the car ( crossing the photo with a shutter time = 1/100 sec. ) the car will be blurred on the photo. Now you take two sheets of paper, draw a dot on one of them and a line on the other ( blurred dot ). The line must exactly be as long as the car has been moving on the photo. Also their moving angle must be the same. FT the dot-picture to D, FT the line-picture to L, FT the photo of the car to C. Then:

IFT( C * ( D / L ) ) and you will have a photo of the car, where you can read its registration number.

Hough transforms are used to recognize lines, circles, parabolas and other mathematical shapes. If such a "known" shape occurs in some photo, the Hough transform will find it and will determine its exact size and location within 1/10 of a pixel-distance. Having a "standard-length" as well in the picture, a computer can calculate very accurate dimension in the picture, check "ovality?" as of things meant to be circular, and so on.

Remember that working machines are often moving very fast, thus the human eye sees nothing. A camera needs perhaps 2μs, using a stroboscope, to see everything in the picture ( well, at least after the computer has calculated for another 100ms ).

 

[dipole]

I don't know why people link to terrible news websites that don't cover the story well - why not just take screenshots of your friends facebook comments and submit that as a news story?

This is the original source.

http://googleresearch.blogspot.co.uk/2015/06/inceptionism-going-deeper-into-neural.html

 

[zoobyshoe]

we just start with an existing image and give it to our neural net. We ask the network: “Whatever you see there, I want more of it!” This creates a feedback loop: if a cloud looks a little bit like a bird, the network will make it look more like a bird. This in turn will make the network recognize the bird even more strongly on the next pass and so forth, until a highly detailed bird appears, seemingly out of nowhere.

What intrigues me about these images is that they could be used to illustrate similar human psychological and neurological flaws, from confirmation bias, through delusion and pareidolia, all the way to hallucination. In fact, I wonder if the exact same kind of feedback loop isn't at work in all those things.